The authors propose a novel approach for open-world semantic segmentation that can accurately identify new categories without additional training data, providing a similarity measure to known categories.
